I am not going to do another technical book [1], but if I did, high on the list would be something like "Software Design by Example" that builds working models of tools for reproducible data science/ML from scratch. Dozens or hundreds of such tools have been built in the last 20 years, and an equal number of papers published, but we seem no closer to converging on a standard model than we were in 2003.
[1] Unless some enlightened company hires me to write.